3. Results
All following performance indexes were calculated from
the validation data sets. Table 3 highlights the average prediction
ability of the different estimators tested (FFANN
and RANN). The calculated errors, RMSE and MAE,
are the average values after applying the 10 calibrated
models of each estimator to both validation data sets. This
table shows that pre-processing improved accuracy of the
soft-sensors by more than 30% in both industrial and
lab-scale distillations. On average FFANN performed
better than RANN when pre-processing is applied.
Table 4 shows the soft-sensor performances of the models
that achieved the smallest RMSE for the validation data
sets, while Figs. 1 and 2 compare model predictions with
their respective measurements. In lab scale distillations,
as data refer to automatic control tuning experiments, the
variability in concentration was larger than routinely
observed under normal operation, especially in run 1
(Fig. 1a). The soft-sensors were highly accurate and,